AI Finds Security Vulnerabilities: What the PHP Warning Means for Mid-Sized Businesses

On October 21, 2026, the BSI (Germany’s Federal Office for Information Security) warned of two critical vulnerabilities in PHP (CVE-2026-109618 and CVE-2026-109619). What you need to do right away is covered in our post on the BSI warning. This article is about what comes after.

Our assessment: this warning was probably just the beginning. AI tools are now finding bugs in code that humans and classic testing tools have reviewed for years. For your company, that means patching has to become part of day-to-day operations rather than an exception.

Table of contents

No reason for PHP bashing

PHP is not a neglected project. It has its own security team, and in 2024 it underwent an external security audit by Quarkslab. Even so, the two vulnerabilities sat in the code for years. Calif names PHP 5.4.0 as the earliest affected version. According to the PHP Foundation, they lie in memory management and error handling, that is, in parts that touch almost every request.

This is not a criticism of PHP. Every large software project has places like these. PHP is simply the first example with a large reach.

What is documented about the PHP vulnerabilities

  • The BSI states in its warning that Wiz and Calif, together with OpenAI, were involved in the discovery.
  • According to OpenAI, Calif found a critical vulnerability with GPT-5.6 Sol. Wiz discovered both vulnerabilities independently and names Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol as the tools used.
  • Calif reports that the first findings came just a few days after the review started on July 30. The target had previously been examined by many people with AI assistance. Calif says it did not need special access or non-public models for this.

An honest assessment belongs here too: part of the problem was not entirely unknown. An out-of-memory bug appeared in 2020 as a seemingly harmless bug report and came up again in the 2024 audit. No attack path had been demonstrated, so nothing was fixed. That this could turn into unauthenticated remote code execution was only shown by the 2026 reviews.

Humans remain indispensable nonetheless. Calif stresses that prior knowledge of the target was decisive. At PHP, according to the PHP Foundation, a human makes every decision and writes or reviews every patch.

Not an isolated case: Firefox, SQLite and AIxCC

PHP is not alone. Three examples with figures from the original sources:

  • Firefox (Mozilla, March 6, 2026): Anthropic’s Frontier Red Team used Claude to find 14 high-severity bugs. In total, these led to 22 CVEs, plus 90 additional bugs. Mozilla counts Firefox among the most thoroughly reviewed codebases on the web. According to Mozilla, the model also uncovered kinds of logic errors that fuzzing had not found until then.
  • SQLite (Google, 2024): The AI agent Big Sleep found an exploitable bug that neither the project’s tests nor OSS-Fuzz had detected. The bug was fixed before the official release. According to Project Zero, Google considers this the first public example of its kind.
  • DARPA AIxCC (2025): In the final, the AI systems found 54 of 63 deliberately planted vulnerabilities and patched 43 of them. They also discovered 18 real, previously unknown vulnerabilities in open source projects (DARPA).

This does not prove that AI is generally smarter than humans. It proves something more specific: when searching for bugs in large codebases, AI-assisted methods have become effective enough to score hits in well-reviewed code that humans and classic tools missed.

Why this will not stop at PHP

Mozilla suspects a considerable backlog of bugs that are only now becoming findable. That affects widely used software. OpenAI expects such findings to increase sharply in the coming months. The PHP Foundation writes that tools find bugs faster than ever, for defenders and attackers alike.

We draw a clear conclusion: expect more reports. They will affect frameworks, shop and CMS extensions, libraries, server software and custom applications. More reports do not mean software is getting worse. The search is getting better. Switching technologies does not solve this, because the next technology also needs maintenance.

The pace is changing, too. On October 21, the BSI considered large-scale exploitation within a few hours of the updates likely, because attackers can analyze patches with AI-assisted tools. On October 22, it was not yet aware of any successful exploitation at scale.

Patching belongs in operations

NIST describes patching as preventive maintenance and as part of the cost of doing business. That is exactly how you should plan for it.

An example: a report affects a library in the customer portal. Nobody knows which version is running. The host points to the development team, and the development team waits for a sign-off. What is missing here is not a faster programmer. What is missing is visibility and ownership. More on this in our post on operational readiness.

Step 1: Inventory applications and dependencies

Create a central application list. For each application, include the business purpose, the responsible person, the development partner, the hosting, key components with versions and the support status. Do not forget old internal tools. For applications that have been running unnoticed for years, the overview is often missing.

Step 2: Define responsibilities

Name a role and a deputy for each task: assessing reports, checking whether you are affected, applying updates, testing, approving. Management clarifies the budget and escalation rules. Hosts, developers and plugin vendors do not automatically take on the same updates. What hosting and technical operations cover must be written into the contract.

Step 3: Prioritize by risk

Not every report is an emergency. Check whether your version is affected, whether the function is reachable and whether there are signs of active exploitation. Ask which data or processes would be at risk and whether a patch exists. Then decide whether the next maintenance window is enough. For critical vulnerabilities in publicly accessible systems, as with PHP now, it usually is not.

Step 4: Make updates repeatable

A reliable process has fixed stages: check whether you are affected, prepare the change, test, secure a recovery path, roll out, then verify function and security status. For a web application, a login, one central business action and one interface are often enough as a test. Plan fixed maintenance windows and keep capacity free for urgent cases. A backup is no substitute for a patch.

Which operating model fits?

In-house operations work if expertise, access, deputies and time are available. An external partner can look after clearly defined layers. For that, they need access, test options and agreed response paths. You can also split the work: the partner handles the technical implementation, and you review the business side internally and give the approval. Put the handovers in writing.

At EXORD, we combine custom software development with technical operations. Which maintenance tasks we take on, we agree on specifically and do not assume implicitly.

Questions for your development partner

In your next conversation, clarify which components your partner monitors and who assesses new reports. Ask which updates are covered by the contract and who adapts the custom application code. Discuss how urgent measures are approved, how the production update status is documented and who steps in during absences. A blanket assurance sounds friendly and leaves the decisive questions open.

Conclusion: preparation instead of technology bashing

AI has found bugs in PHP, Firefox and SQLite that stayed undiscovered for a long time. We expect many more projects to follow in the coming months. Anyone who then does not know which software they run and who updates it will lose time they do not have.

Start with the application list. Then define responsibilities including deputies, and walk through the update process with the next change that is due. If you lack capacity or clarity internally, talk to us in IT consulting about a review of your applications and maintenance responsibilities.

Frequently asked questions about AI and security vulnerabilities

Were the PHP vulnerabilities found with AI?

Yes. According to OpenAI, Calif found one of the vulnerabilities with GPT-5.6 Sol. Wiz discovered both vulnerabilities independently and also names AI models as tools. The BSI lists Wiz and Calif, together with OpenAI, as involved parties. Humans assessed the findings and wrote the patches.

Is PHP now less secure than other technologies?

No. The warning shows that AI can find bugs in well-maintained code. That applies to every large piece of software. What matters are current versions and a practiced update process.

Is updating my application enough?

Not for these vulnerabilities. The flaws are in PHP itself, so the PHP version on the server must be patched. The affected and patched versions (8.2.35, 8.3.36, 8.4.27 and 8.5.12) are listed in the advisories for CVE-2026-109618 and CVE-2026-109619.

Sources

  1. BSI: PHP – Kritische Schwachstellen bedrohen Webanwendungen weltweit (security warning in German, version 1.2, October 22, 2026)
  2. BSI: PHP – BSI warnt vor kritischen Schwachstellen, Sicherheitsupdate noch heute notwendig (press release in German, October 21, 2026)
  3. The PHP Foundation: Today’s PHP Security Release, and How We’re Strengthening PHP Security (October 21, 2026)
  4. The PHP Foundation: How the PHP Security Team Handles Your Reports (October 14, 2026)
  5. php-src Security Advisory: RCE via OOM corruption (CVE-2026-109618, GHSA-w8hw-w9fx-j79w)
  6. php-src Security Advisory: RCE via stale global state (CVE-2026-109619, GHSA-6wmf-4rv5-p3rp)
  7. Calif Research: PHP2Shell, A tale of two PHP remote vulnerabilities
  8. Wiz Blog: PHP2Shell, AI uncovers critical RCE vulnerabilities (October 21, 2026)
  9. OpenAI: Patch the Planet, global response coordination as agents uncover critical vulnerabilities (October 21, 2026)
  10. The Mozilla Blog: Hardening Firefox with Anthropic’s Red Team (March 6, 2026)
  11. Google Project Zero: From Naptime to Big Sleep, Using Large Language Models To Catch Vulnerabilities In Real-World Code (November 1, 2024)
  12. DARPA: AI Cyber Challenge marks pivotal inflection point for cyber defense (2025)
  13. NIST SP 800-40 Rev. 4: Guide to Enterprise Patch Management Planning, Preventive Maintenance for Technology (04/2022)